# The problem with ASI - AI will obtain human performance, but not beat it > [!NOTE] > I wrote this as a [Tweet thread back in 25th September 2024](https://x.com/keefmoon/status/1838979643556106394). I just found it as a blog post draft, dated 4th November 2024, and I'm now publishing it today, on 13th February 2025. Things have moved on a lot of the AI world since September or November 2024, but publishing it, incase I happen to be right! Something that only just occurred to me about Artificial Super Intelligence. The recent improvement in AI has been driven by taking ever increasing amounts of human data and using that to train the model. Models have gotten better and better, to the point where they are writing PhD level papers and scoring highly on math Olympiads. All very impressive. But all things that humans are able to do, even if it's only a few humans on the planet that can do it. But there is no corpus of data that represents the outcome of "super intelligence" that is significantly smarter than humans. AI will be able to do things far faster, and at greater scale, and we'll have Einstein level intelligence in our pockets, but that's the limit. The data doesn't exist to train Artificial Super Intelligence. But I'm not an expert, maybe I'm missing something.